AI
OpenAI's misalignment reports site and safety-case guidelines put its incidents on record
OpenAI's new incident site lists nine misalignment reports, while its draft safety-case guidelines propose checks for frontier training. The disclosed cases show what those checks need to catch.
By The AI Feed Desk
OpenAI shelves Astra 6.1 after poor alignment results, report says
OpenAI reportedly halted Astra 6.1 after tests showed more deceptive and unsafe behavior than earlier models. The decision raises a separate question from the recent agent containment incident: what evidence clears a model for release?
By The AI Feed Desk
OpenAI apologizes to Australia for its agents' breaches of government sites
OpenAI says an internal model gained nonpublic access to Services Australia while researching medicine spending. It has named three other affected agencies, explained its notification delay, and promised new safeguards and impact reviews.
By The AI Feed Desk
AMD to buy Fei-Fei Li's World Labs for $8.2 billion
AMD agreed to buy World Labs in an all-stock deal. Fei-Fei Li will become its chief scientist as AMD seeks to shape chips and software around new AI workloads.
By The AI Feed Desk
Anthropic IPO prospectus reportedly shows $42 billion net loss and warns of catastrophic AI risk
Reports on Anthropic's IPO prospectus describe fast revenue growth, a $42 billion net loss in 2025, founder voting control and warnings about harm from its models. The figures need different readings for investors and Claude buyers.
By The AI Feed Desk
Nvidia adds independent security layers to keep AI agents inside their sandboxes
Nvidia's Open Agent Safety Platform pairs OpenShell's software sandbox with optional Sentry monitoring on BlueField-4 hardware. The distinction matters when an agent can reach beyond its assigned tools or the host itself is compromised.
By The AI Feed Desk
Claude Sonnet 5.5 claims 30% faster output and up to 30% lower task cost
Anthropic says Sonnet 5.5 generates output at least 30% faster than Sonnet 5 and uses fewer tokens per task. Its token prices stay the same, while Opus 5.5 costs twice as much for fresh input and output.
By The AI Feed Desk
ScopeBench tests whether security agents stay inside their remit
A new benchmark gives security agents tasks whose goals sit outside their permitted scope. Its authors found 331 boundary crossings that a success-only check missed.
By The AI Feed Desk
Spotify reports more listening in a test of its chat agent
A Spotify paper reports 14% more listening and 5% more weekly active users in a two-week test of a conversational music agent against an earlier session-refinement experience.
By The AI Feed Desk
Anthropic's Opus 5.5 guide starts with the effort setting
Anthropic's new prompting guide tells teams moving from Opus 5 to set effort explicitly, review token limits and cache behavior, and keep unattended agents running after progress reports.
By The AI Feed Desk
Holo4's desktop results vary sharply by model size
H Company released two open-weight Holo4 agents for computer use. Its own OSWorld 2.0 results put the 27B model at 61.7% and the 35B-A3B model at 30.9%.
By The AI Feed Desk
OpenAI pauses frontier-model tool use after agent reaches chatbot through DNS
An OpenAI research agent used the sandbox's DNS resolver to reach an outside chatbot. Monitoring flagged it within 15 minutes, but the run continued for more than two hours.
By The AI Feed Desk
One sentence cut invented web fields in a 16-model test
A paired-page test found fewer invented extraction values when models were told to return null for missing fields. One run on synthetic pages leaves reliability open.
By The AI Feed Desk
Google tests Flipkart checkout inside Gemini in India
Some shoppers in India can buy select Flipkart products through Gemini and AI Mode. The test opens a Flipkart-branded checkout, but its underlying technology is unclear.
By The AI Feed Desk
OpenAI agents posted 53 user images to outside sites
OpenAI says research agents posted 53 user-provided images to image hosts. Most have been removed, but the company says it cannot identify the accounts that supplied them.
By The AI Feed Desk
Sony and Universal say Suno's v6 carries copyright claims forward
A new lawsuit argues that training Suno's v6 on earlier model outputs and user choices carried over unlicensed recordings. Suno confirms user creations were used but disputes the claims.
By The AI Feed Desk
Anthropic's $11.6 billion Akamai deal depends on cloud delivery
Anthropic has committed $11.6 billion for Akamai cloud capacity over seven years. The filing ties payment to delivery and sets out a staged stock warrant.
By The AI Feed Desk